Skip to main content

What Is an Agent Native Database

9 min read

As large language models continue to advance, AI application architectures are undergoing a fundamental transformation.

In the past, we focused on how to “feed” external data to models. Now, as Agents become the main actors in applications, we need to ask: what kind of data foundation is needed to support their long-term operation and continuous evolution?

That is the motivation behind RAGFlow 1.0’s new positioning: evolving from an end-to-end RAG engine into an Agent Native Database. Today, with the official release of RAGFlow 1.0, the Agent Native Database has moved from concept to implementation.

I. Architectural Shift: Why Must the Context Layer Be Independent?

To understand the Agent Native Database, we first need to clarify the current evolution of the Agent ecosystem.

Over the past year, components of the Agent Harness—including task orchestration, context engineering, Skills systems, and memory operations—have increasingly become standardized and production-ready. Meanwhile, the Managed Agents concept proposed by Anthropic earlier this year offers a highly instructive architectural abstraction: LLM and Harness are treated together as the “Brain”—a stateless decision core responsible for planning, reasoning, and dispatch. An Agent's actual capabilities are determined by the tools it can invoke.

This is a critical architectural dividing line: the Brain remains stateless and can be started or stopped as needed, while all persistent state is stored externally. The entire ecosystem is thus moving toward a clear separation between the “state layer” and the “stateless layer.” If the Harness were made responsible for both reasoning and heavy data persistence, file parsing, and state rollback, the system would become not only excessively complex but also difficult to scale.

Based on this evolution, a three-layer understanding of Agent architecture has become consensus:

  1. Model layer: responsible for understanding, reasoning, and decision-making.
  2. Harness layer: responsible for loop control, task planning, Skills invocation, and read/write control of memory.
  3. Context layer: responsible for storage, processing, and retrieval of persistent state.

RAGFlow’s positioning is to focus on the outermost Context layer. We are committed to processing data and supplying context so that the Harness above can operate with a light footprint.

II. Why Can’t Existing Databases Simply Be “Extended”?

When building an Agent data foundation, the most natural approach in the industry is to extend existing data platforms, such as data lakes and data warehouses, with vector search and full-text search capabilities. This approach inherits directly from existing experience with big data platforms: it seems that installing a vector plugin is enough to form the prototype of an “all-purpose foundation.”

However, when we apply this approach to Agent scenarios, we quickly hit a wall:

First, processing unstructured documents goes far beyond storage and retrieval. An Agent’s requirements for documents and other unstructured data are not limited to “can store it, can find it.” They require a complex, end-to-end pipeline for ingesting, governing, cleaning, and semantically enriching data from heterogeneous sources. From complex PDF layout parsing and table structure reconstruction to using large models to extract entities, relations, and claims, and then building a navigation tree with evidence chains—this is fundamentally heavy data engineering, and it is difficult to accomplish elegantly as a black box inside a closed database kernel.

Second, retrieval demands dynamic and sophisticated capabilities. When an Agent invokes retrieval, which indexing strategy should be used? Does retrieval require additional context completion or navigation? Should the retrieved results return the full original text to the LLM, or only high-density fragments? This logic is highly complex and dynamic, cannot be expressed by traditional database query languages, and must be provided by a tool layer closer to the Agent.

Third, structured data systems do not need to be rebuilt. Enterprises’ existing business systems already carry a large amount of structured assets; covering them again with a new database is neither realistic nor necessary. What the Agent really needs is an access interface to these systems, along with supporting documentation. As long as the foundation can provide feature-rich search over this “document-like metadata,” it is enough to help the Agent efficiently filter out tool options that do not match the current intent.

Taken together, an Agent’s data foundation absolutely cannot be just “a database with retrieval capabilities.” So what should it be?

The so-called Agent Native Database is not an existing database with a vector plugin bolted on. It is a sophisticated context data control plane built around Retrieval, with large models as the processing engine and Agents as the retrieval consumers.

III. Retrieval: A Data Control System Spanning Strategy and Engineering

Since it is a data control plane, what capability matters most in the Context layer? The answer is undoubtedly Retrieval.

But what supports Retrieval is far more than a simple search interface. RAGFlow’s ability to provide powerful retrieval services in the Agent era rests on two underlying infrastructures: the Ingestion Pipeline and the knowledge compilation system. Together, they form the foundation of RAGFlow's competitive advantage in Retrieval.

1. Ingestion Pipeline: An Enterprise-Grade Data Governance Foundation

Highly available retrieval depends first and foremost on reliable data ingestion. The RAGFlow 1.0 Pipeline provides enterprise-grade governance mechanisms—the kind of “thickness” that many lightweight RAG solutions find hard to match:

  • Fault tolerance guarantees: if processing is interrupted at any stage, a document is either fully ingested or not ingested at all. Task-level redelivery, segment-level isolation, and idempotent overwrites ensure that data entering the retrieval system never leaves behind “dirty data” or a “half-written document.”
  • Versioning and metadata: through two parallel versioning tracks—“execution orchestration” and “knowledge provenance”—along with strict channel isolation and lineage tracking, every piece of evidence retrieved can be traced to an exact version and original text, making the Agent’s data supply absolutely trustworthy.

2. Knowledge Compilation: A New Paradigm for LLM-Managed Data

On top of a reliable Pipeline, Knowledge Compilation not only reshapes the economics of token usage but also represents a new paradigm for reconstructing data management with LLMs. It is by no means merely for generating summaries; it is a customizable, extensible template engine:

  • Multiple forms of knowledge artifacts: The system currently supports tree-based navigation (providing hierarchical localization), graph exploration (providing entity-relation association), and LLM Wiki (providing synthesized reading), among other forms. Going forward, it will continue to expand with Ontology and Memory templates.
  • Constraining and guiding retrieval: Compilation gives knowledge a semantic structure. This enables the Harness to retrieve information more purposefully, rather than relying on blind trial and error, by clearly knowing “within what scope to search and what type of thing to look for.”
  • Extreme Token compression: it transforms long text into high-density “claim–evidence pairs.” On a hit, a single query can save more than 90% of Tokens; on a miss, it adds only about 2% consumption. This enables Agents to accomplish more within a limited context window.

IV. The Interface Between Context and Harness: Minimal Tool Primitives

Precisely because the underlying Pipeline and knowledge compilation system push “governance” and “compression” to the extreme, RAGFlow can keep its external interface minimal.

The guiding principle of infrastructure design is simple: the more complexity handled internally, the simpler the external interface should be.

The Harness does not need to care about the complex parsing and index construction behind it; it only needs to interact through the most standard primitives:

  • search / grep: lexical or semantic recall to probe for clues.
  • read: read the original text or evidence by anchor to verify details.
  • navigate / explorer: drill down in the navigation tree or graph to explore relationships.

On each call, RAGFlow returns only candidates and evidence, with precise citations and metadata. How to assemble these results into the Context Window and how to control the Token budget is entirely up to the Brain in the Harness. The Harness decides what it needs; RAGFlow delivers the right evidence efficiently and with minimal token overhead. This clear contract allows any Agent to obtain the strongest retrieval capability at extremely low integration cost.

V. Dual Form: Built-in Harness and Full Ecosystem Embrace

In actual product deployment, RAGFlow 1.0 plays both the role of “reference implementation” and “enabling platform.”

On the one hand, RAGFlow comes with a powerful built-in Harness. Not every team has the resources to develop a complex Agent orchestration framework from scratch. RAGFlow includes a carefully tuned Harness that distills complex loop control into standard flows (such as evidence prefetching, the evidence gate, and the tool loop). Developers do not need to manage the underlying execution loops; RAGFlow delivers a fast, out-of-the-box experience.

On the other hand, RAGFlow will fully embrace the external Harness ecosystem. Many large enterprises already have their own orchestration layers, such as complex workflows built on LangGraph or fully self-developed Agent systems. RAGFlow will not require teams to replace their existing orchestration frameworks. Instead, it aims to serve as a powerful extension to external Harnesses.

To streamline integration with external Harnesses, RAGFlow plans to introduce a comprehensive CLI (command-line interface) integration solution. Regardless of the language in which your Harness is written, you can seamlessly access its powerful parsing, compilation, and retrieval tools by calling RAGFlow’s CLI or standardized API. This allows RAGFlow to function like a local daemon, providing high-quality context to your Agent whenever needed.

VI. Conclusion: From RAG Engine to the Cornerstone of Agents

Moving from an end-to-end RAG engine to an Agent Native Database, the evolution of RAGFlow 1.0 reflects a rethinking of the relationship between data and Agents.

As the Brain becomes stateless and the Harness becomes standardized, the Context layer—which carries all state and memory—becomes the key infrastructure determining whether Agents can truly move into production.

The Agent Native Database is the foundation born to meet this need. It handles the complexity of unstructured data and delivers clean, high-density knowledge to Agents.

Today, the release of RAGFlow 1.0 officially opens this architectural evolution. With a powerful built-in Harness and through APIs and the forthcoming CLI, it embraces all Harnesses. In an era of accelerating Agent explosion, only by making the context supply layer deep and thorough can Agents move beyond their early-stage limitations and become productivity tools that enterprises can trust.

And RAGFlow stands at the beginning of this new journey.